Hugging Face Overview
Sungur-14B: Turkish-Specialized Reasoning Model
Sungur-14B is a 14 billion parameter large language model developed by suayptalha, built upon the Qwen/Qwen3-14B architecture. Its primary distinction lies in its specialization for the Turkish language, particularly in reasoning tasks.
Key Capabilities & Training
- Turkish Reasoning Focus: The model was fine-tuned using
suayptalha/Sungur-Dataset, a collection of 41.1k Turkish reasoning-focused conversations covering mathematics, medicine, and general knowledge. - Efficient Fine-Tuning: Training utilized 4-bit QLoRA for Supervised Fine-Tuning (SFT), preserving the base model's capabilities while adapting it for Turkish.
- Thinking Mode: Sungur-14B includes an "enable_thinking" feature, allowing the model to engage its reasoning abilities to enhance response quality, especially for complex problems like mathematical equations.
Performance & Benchmarks
Sungur-14B demonstrates competitive performance in Turkish benchmarks, often outperforming its base model, Qwen3-14B, in specific metrics:
- ARC (tr, acc): 0.4727 (vs 0.4701 for Qwen3-14B)
- HellaSwag (tr, acc): 0.4051 (vs 0.4017 for Qwen3-14B)
- Winogrande (tr): 0.5893 (vs 0.5656 for Qwen3-14B)
- Turkish GSM8K: Achieves 77.60 (strict), placing it competitively among larger models like Qwen2.5-32B-Instruct and surpassing Llama-3-1-70B-Instruct in this specific Turkish math benchmark.
Ideal Use Cases
- Turkish Analytical Dialogue: Generating logically structured and context-aware responses in Turkish.
- Education: Assisting with problem-solving and explanations in Turkish educational contexts.
- Domain-Specific Problem Solving: Applications requiring strong native Turkish reasoning in fields like mathematics and medicine.